In the past year, a dominant theme in the Artificial Intelligence conversation has been its “black box” problem. MIT Technology Review called it “The Dark Secret at the Heart of AI.” In defense of machine learning, staunch supporters of the technology have argued that it is no less transparent than the human mind. That is, they are both opaque. These arguments miss the point. Instead of comparing the performance of algorithms to the performance of the human mind, we should evaluate the performance of algorithms vis-a-vis the requirement of the person impacted by it.
In a New York Times Op-Ed, Vijay Pande, Ph.D., a General Partner at Andreessen Horowitz, a venture capital firm, argues that “Artificial Intelligence’s ‘Black Box’ Is Nothing to Fear.” Using a doctor visit as an example, Dr. Pande argues that when the patient asks the doctor how she made the diagnosis, she would “probably share some of the data she used to draw her conclusion” but she could not “really explain how she made that decision.” What he fails to mention is that machine learning could not even begin to explain the diagnosis. And therein lies the problem.
Imagine that you are sick. You do not have a general practitioner that you typically go to, so you pick two doctors online that you want to try out. You go to the 1st doctor, and she tells you that you have the flu. You ask her why she thinks that and she then explains how the influenza virus impacts the body and how that impact has surfaced as the symptoms that you are experiencing. You ask her if she is 100% sure, and she says “no,” but she is very confident that it is the flu and if things do not improve in 48-72 hours, you should come back to the see her. You then go to the 2nd doctor, and he also tells you that you have the flu. You ask him why he thinks that and he looks at you with a blank stare. That is it. You walked into the room, said “I have symptoms X, Y, Z” and the doctor blurted out “flu,” and then you left. Regardless if both were right, out of these two doctors, which one would you trust and which one would you visit again?
The fact of the matter is that we as humans want to know “why,” especially when someone or something makes a decision that impacts us. Delivering on the promise of AI is less about building algorithms that precisely simulate or emulate the human brain and more about developing algorithms as tools that work seamlessly within the framework of everyday life. A good example of this fundamental product design principle comes from Dr. Pande’s partner, Marc Andreessen. When Tim Berners-Lee criticized him for enabling multi-media capabilities in Mosaic, the first widely used web browser, Mr. Andreessen replied “If people want images, they get images. Bring it on.”
Technology is supposed to conform to us, not the other way round.
Founder and CEO